Università degli Studi del Molise Facoltà di Economia Dipartimento di Scienze Economiche, Gestionali e Sociali Via De Sanctis, I-86100 Campobasso (Italy) ECONOMICS & STATISTICS DISCUSSION PAPER No. 55/09 A Characterization of the Dickey-Fuller Distribution, With Some Extensions to the Multivariate Case by Roy Cerqueti University of Macerata, Dept. of Economic and Financial Institutions Mauro Costantini University of Vienna, Dept. of Economics and Claudio Lupi University of Molise, Dept. SEGeS The Economics & Statistics Discussion Papers are preliminary materials circulated to stimulate discussion and critical comment. The views expressed in the papers are solely the responsibility of the authors. A Characterization of the Dickey-Fuller Distribution, With Some Extensions to the Multivariate Case Roy Cerqueti∗ Mauro Costantini† Claudio Lupi‡ Abstract This paper provides a theoretical functional representation of the density function related to the DickeyFuller random variable. The approach is extended to cover the multivariate case in two special frameworks: the independence and the perfect correlation of the series. key words: Dickey-Fuller distribution, unit root JEL codes: C12, C16, C22 1 Introduction In this paper we deal with the theoretical case where n possibly cross-dependent time series are generated by yi,t = yi,t−1 + ui,t = yi,0 + t X (i = 1, . . . , n ; t = 1, . . . , T ) ui,j j=1 = yi,0 + Si,t (1) where yi,0 = ci with probability one, or it has a given probability distribution. The ui,t ’s are assumed to satisfy some regularity conditions so that a suitably normalized transform of Si,t , ∗ (r) := T −1/2 σ −1 S Si,T i,bT rc (where b·c denotes the integer part and r ∈ [0, 1]), is such that i ∗ Si,T (r) ⇒ Wi (r) as T → ∞, with Wi (r) a Wiener process (see e.g., Phillips, 1987, for a detailed discussion of such regularity conditions and of the exact meaning of the normalization). Here we take the simplifying assumption that ui,t ∼ iid(0, σi2 ), so that the conditions for the weak convergence of the normalized partial sums are trivially satisfied. It is well known that, under the null H0 : ρi = 1, the t-ratio based on the OLS estimator of ρi , b tρi , in ∆yi,t = ρi yi,t−1 + ei,t . (2) ∗ (corresponding author) University of Macerata, Dept. of Economic and Financial Institutions, Via Crescimbeni, 20. I-62100 Macerata (Italy). roy.cerqueti@unimc.it † University of Vienna, Dept. of Economics, Bruenner Strasse, 72. A-1210 Vienna (Austria). mauro.costantini@univie.ac.at ‡ University of Molise, Dept. of Economics, Management and Social Sciences, Via De Sanctis. I-86100 Campobasso (Italy). lupi@unimol.it 1 2 The Univariate Case 2 has the non-standard Dickey-Fuller limiting distribution 1 Z b tρi ⇒ Z Wi (r)dWi (r) 0 1 Wi2 (r)dr − 21 (3) 0 as T → ∞ (see e.g., Phillips, 1987). Building on Ruben (1962), in Section 2 we derive an explicit formulation of the conventional univariate Dickey-Fuller distribution (3). Although not easy to manage, the proposed formulation can be used to derive analytical results, as opposed to the conventional simulation approach. In this respect, our work is related to Abadir (1995). In Section 3 we extend the analysis to the multivariate case by focusing on the special cases where the series are either independent or perfectly correlated. Section 4 concludes. In the derivation of the theoretical results, we assume that all the random quantities introduced in the paper are contained in a filtered probability space (Ω, F, {Ft }t>0 , P ). 2 The Univariate Case Consider the Dickey-Fuller distribution (3): by definition of stochastic integral, we take a partition of the interval [0, 1] in N intervals of length 1/N . If N is large enough, we can approximate the asymptotic distribution of b tρi under the null as follows: Z 1 Z Wi (r)dWi (r) 0 ∼ 1 Wi2 (r)dr − 12 ∼ 0 N X ! Wi (k/N ) [Wi (k/N ) − Wi ((k − 1)/N )] k=1 k=1 = 1 Wi2 (1) − 1 2 !− 12 N 1 X 2 Wi (k/N ) = N 1 N N X !− 12 Wi2 (k/N ) =: Xi /Yi , (4) k=1 where Xi := Yi := 1 Wi2 (1) − 1 2 N 1 X 2 Wi (k/N ) N (5) ! 21 . (6) k=1 Therefore, by definition of Wiener process, the distribution of the univariate Dickey-Fuller test under the stated conditions can be approximated as follows: 1 2 (1) − 1 χ b tρi ⇒ Xi /Yi ∼ 2 (7) 1 , (CvM0 (1)) 2 where χ2 (1) denotes a Chi-squared distribution with 1 degree of freedom and CvM0 (1) denotes a zero level Cramér-von Mises distribution with 1 degree of freedom. 2 The Univariate Case 3 In order to derive an analytical expression for the density function of the Dickey-Fuller distribution, we need to introduce the ratio variable Zi := Xi /Yi . The cumulative distribution function of Zi will be denoted as Fi . For z ∈ R, we have Fi (z) = P (Zi ≤ z) = P (Xi /Yi ≤ z) Z +∞ Z yz fXi ,Yi (x, y)dxdy , = 0 (8) 0 where fXi ,Yi is the joint density function of the random variables Xi and Yi . The density function fi of the variable Zi is Z +∞ Z yz ∂ fi (z) = fXi ,Yi (x, y)dxdy ∂z 0 0 Z +∞ yfXi ,Yi (yz, y)dy . = (9) 0 We need to find an explicit form of the density function fXi ,Yi , in order to get an expression for fi . Fixed x, y ∈ R, we have fXi ,Yi (x, y) = fYi |Xi (y|Xi = x)fXi (x) , (10) where fYi |Xi is the density function of Yi conditional on Xi , and fXi is the (marginal) density function of the random variable Xi . We write the cumulative distribution functions of Xi as: FXi (x) = P (Xi ≤ x) 1 2 = P (W (1) − 1) ≤ x 2 i = P (Wi2 (1) ≤ 2x + 1) Z 2x+1 = KXi s−1/2 e−s/2 ds , (11) −∞ where KXi is the normalizing constant. The density function fXi is then Z 2x+1 ∂ fXi (x) = KXi s−1/2 e−s/2 ds ∂x −∞ 2x + 1 −1/2 . = 2KXi (2x + 1) exp 2 (12) 2 The Univariate Case 4 The cumulative distribution function of Yi conditional on Xi = x is FYi |Xi (y|Xi = x) = P (Yi ≤ y|Xi = x) !1/2 N X 1 = P Wi2 (k/N ) ≤ y Wi2 (1) = 2x + 1 N k=1 "N −1 #!1/2 X 1 = P Wi2 (k/N ) + 2x + 1 ≤ y N k=1 ! N −1 X 2 2 = P Wi (k/N ) + 2x + 1 ≤ N y = P k=1 N −1 X ! Wi2 (k/N ) ≤ N y 2 − 2x − 1 . (13) k=1 Equation (13) suggests that we need to discuss the distribution of a sum of squared non-independent zero-mean Gaussian variables. Denote the Cramér-von Mises distribution as V := N −1 X Wi2 (k/N ) , (14) k=1 and FV (v) = P (V ≤ v) , v ∈ R+ . (15) Our approach relies on an invariant symmetry property of the random V (see Ruben, 1962). In particular, Ruben (1962) shows that the cumulative distribution function of V can be written as series expansions of Chi-squared cumulative distribution functions, i.e. there exists a sequence of real numbers {λj }j∈N such that FV (v) = +∞ X λj FN −1+2j (v) , (16) j=0 where FN −1+2j is the cumulative distribution function of a Chi-squared random variable with (N − 1 + 2j) degrees of freedom. By substituting the explicit expression of the F ’s in (16), we obtain Z w +∞ X FV (v) = λj e−s/2 s(N −3+2j)/2 ds , (17) j=0 0 where we assume without loss of generality that the λ’s contain also the normalizing constants related to the Chi-squared distributions. By (17), then (13) can be written as ! N −1 X FYi |Xi (y|Xi = x) = P Wi2 (k/N ) ≤ N y 2 − 2x − 1 k=1 = +∞ X j=0 Z λj 0 N y 2 −2x−1 e−s/2 s(N −3+2j)/2 ds . (18) 3 The Multivariate Dickey-Fuller Distribution 5 The conditional density function fYi |Xi is +∞ X fYi |Xi (y|Xi = x) = j=0 ∂ λj ∂y "Z # N y 2 −2x−1 −s/2 (N −3+2j)/2 e s ds 0 N y 2 − 2x − 1 × = 2N y · exp − 2 +∞ X × λj [N y 2 − 2x − 1](N −3+2j)/2 . (19) j=0 By substituting (12) and (19) into (10), we have y N y 2 − 2(2x − 1) fXi ,Yi (x, y) = 4N K̄Xi √ × · exp − 2 2x + 1 +∞ X × λj [N y 2 − 2x − 1](N −3+2j)/2 . (20) j=0 Finally, from (9) and (20) we have Z fZi (z) = 4N K̄Xi +∞ 0 × +∞ X y2 N y 2 − 2(2yz − 1) √ · exp − × 2 2yz + 1 λj [N y 2 − 2yz − 1](N −3+2j)/2 dy . (21) j=0 Although rather involved, (21) can in principle be used to derive analytical results on the DickeyFuller distribution. 3 The Multivariate Dickey-Fuller Distribution This section is devoted to the analysis of the distribution of the multivariate Dickey-Fuller t-ratio. Let’s define the random vector (Z1 , . . . , Zn ) of asymptotic distributions under the null, accordingly with (3) and (6). More precisely, we can write 1 Wi2 (1) − 1 Zi := 2 !− 21 N 1 X 2 Wi (k/N ) N i = 1, . . . , n . (22) k=1 Our aim is to provide a closed form expression for the joint density function fZ1 ,...,Zn of the random vector (Z1 , . . . , Zn ). Here we study the two extreme cases where the series are either independent or perfectly correlated. 3.1 Independent Series Assume that the series y’s are cross sectional independent. Once that the univariate (marginal) density has been derived, this case becomes trivial. Indeed, for each (z1 , . . . , zn ) ∈ Rn , we can 3 The Multivariate Dickey-Fuller Distribution 6 write the density function fZ1 ,...,Zn simply as fZ1 ,...,Zn (z1 , . . . , zn ) = n Y fZi (zi ) , (23) i=1 where fZi (zi ) is given by (21). 3.2 Perfectly Correlated Series Fixed i = 2, . . . , n, we assume that yi and yi−1 are perfectly correlated, and there exists a constant αi such that yi = αi yi−1 . (24) Of course, the ordering of the series is purely conventional. Any ordering would be possible just by changing the parameter αi . The dependence among the variables is reflected on the dependence among the Wiener processes Wi . In particular, by using the derivation of the Dickey-Fuller asymptotic distribution, then condition (24) can be rewritten in terms of the Wiener processes Wi : Wi = αi Wi−1 . (25) By substituting (25) into (22), we obtain Zi := = 1 2 1 N (αi Wi−1 (1))2 − 1 1 PN 2 2 2 k=1 (αi Wi−1 (k/N ) |αi | Zi−1 + 1 2 2 (αi α2i N − 1) 1 2 (k/N ) 2 W k=1 i−1 i = 1, . . . , n . (26) PN Formula (26) allows us to write explicitly the conditional cumulative distribution of Zi given Zi−1 . Consider zi , zi−1 ∈ R. Then (18) gives P (Zi ≤ zi |Zi−1 = zi−1 ) = 1 2 2 (αi − 1) ≤ z = P |αi |Zi−1 + Zi−1 = zi−1 i 1 2 2 αi P N 2 k=1 Wi−1 (k/N ) N " #2 N X 2|α |(z − |α |z ) i i i i−1 2 . √ =P Wi−1 (k/N ) ≤ 2 − 1) N (α i k=1 (27) Now, consider the joint density function of the random variable (Z1 , . . . , Zn ). Given (z1 , . . . , zn ) ∈ Rn , the dependence condition (25) implies fZ1 ,...,Zn (z1 , . . . , zn ) = fZ1 (z1 ) · n Y i=2 fZi |Zi−1 (zi |Zi−1 = zi−1 ) , (28) 4 Conclusions 7 where ∂ [P (Zi ≤ zi |Zi−1 = zi−1 )] . ∂zi Consider i = 1 . . . , n and zi , zi−1 ∈ R. Then (18) gives » –2 2|αi |(zi −|αi |zi−1 ) Z +∞ √ X N (α2 −1) i P (Zi ≤ zi |Zi−1 = zi−1 ) = λj e−s/2 s(N −1+2j)/2 ds . fZi |Zi−1 (zi |Zi−1 = zi−1 ) = j=0 (29) (30) 0 The density function of Zi conditional on Zi−1 , fZi |Zi−1 , is then » –2 Z 2|αi√|(zi −|αi |zi−1 ) +∞ X ∂ N (α2 −1) i e−s/2 s(N −1+2j)/2 ds fZi |Zi−1 (zi |Zi−1 = zi−1 ) = λj ∂zi 0 j=0 " #2 − |αi |zi−1 ) 1 2|αi |(zi − |αi |zi−1 ) √ = × · exp − 2 N (αi2 − 1)2 N (αi2 − 1) " #N −1+2j +∞ X 2|αi |(zi − |αi |zi−1 ) √ × λj . (31) N (αi2 − 1) j=0 8αi2 (zi 4 Conclusions In this paper an explicit approximation of the density function of the multivariate Dickey-Fuller random variable is provided. We proceed by analyzing at first the univariate case. Our result is grounded on an invariant symmetry property of some random variables involved in the DickeyFuller distribution (see Ruben, 1962). As in Abadir (1995), the followed approach allows us to avoid the conventional simulationbased approach. The theoretical results regarding the univariate case are then extended to the multivariate framework under the assumptions of independent and perfectly correlated series. Although the independent and the perfectly correlated cases are two extreme settings, they represent the starting point for exploring models with less restrictive assumptions. In this respect, the analysis of a general cross sectional dependence case is already in our research agenda. Moreover, while we deal here only with the Dickey-Fuller distribution in the absence of deterministic terms, further extensions are under scrutiny to cope with the “constant” and “constant plus linear trend” cases. References Abadir, K. M. (1995), “The Limiting Distribution of the t Ratio under a Unit Root,” Econometric Theory, 11, 775–793. Phillips, P. C. B. (1987), “Towards a Unified Asymptotic Theory for Autoregression,” Biometrika, 74, 535–547. Ruben, H. (1962), “Probability Content of Regions Under Spherical Normal Distributions, IV: The Distribution of Homogeneous and Non-Homogeneous Quadratic Functions of Normal Variables,” Annals of Mathematical Statistics, 33, 542–570.